Industrial Safety Alarm Event Recording and Tracing Method Based on Graphic Control System

Abstract

The present invention provides an industrial safety alarm event recording and tracing method based on a graphic control system, which relates to the field of industrial automation technology. The method includes: performing undersampling processing on the collected original alarm signals, and using a sparse reconstruction algorithm to generate reconstructed alarm signals; according to the pre-configured area division rules and alarm type classification rules, dividing the reconstructed alarm signals into alarm events of different alarm levels and alarm areas; at the same time, writing the level, area, occurrence time and processing status of the alarm events into the historical database by associating with PLC data points; and after receiving the query instruction from the user, extracting the historical records of the alarm events from the database and performing visual tracing on the graphic control system interface. This method can accurately retain the key features of the alarm signals while reducing the transmission and storage pressure of massive alarm data, realize the automatic recording and rapid tracing of multi-source heterogeneous industrial alarms, and improve the efficiency and reliability of safety alarm management.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial automation, and more particularly, to a method for recording and tracing industrial safety alarm events based on a graphic control system. Background Art

[0002] In the process of large-scale industrial production, factories usually deploy a large number of sensors and PLCs (programmable logic controllers) to obtain and transmit production operation information in real time, and implement industrial alarm monitoring within a management platform (such as a graphic control system). However, with the increasing complexity of industrial production processes, the number of alarm signals has increased sharply, and the sampling periods and data formats in different factory areas or different work sections are not the same, resulting in problems such as large transmission bandwidth pressure, difficulty in quickly storing and analyzing massive data, and difficulty in effectively integrating multi-source information for traditional alarm management solutions. In addition, if only high-frequency sampling under a single working condition is used to capture alarm signals, a large amount of redundant data is often generated, the alarm correlation across regions or factory areas cannot be fully utilized, false alarms or missed alarms are likely to occur, and there is a lack of tools for fine tracing and visual tracking of historical records, bringing great challenges to industrial safety operation and maintenance. Summary of the Invention

[0003] Aiming at the deficiencies of the prior art, the present invention provides a method for recording and tracing industrial safety alarm events based on a graphic control system, including:

[0004] Collecting original alarm signals, performing undersampling processing on the original alarm signals to generate compressed alarm signals; for the compressed alarm signals, reconstructing them through a sparse reconstruction algorithm to generate reconstructed alarm signals;

[0005] According to preset area division rules and alarm type classification rules, dividing the reconstructed alarm signals into alarm events of different alarm levels and different alarm areas;

[0006] According to the divided alarm events, automatically recording the alarm level, area information, occurrence time and processing status of the alarm events into a historical database through pre-associated PLC data points to form a historical record of alarm events;

[0007] Receiving a query instruction from a user, and according to the query instruction, extracting the corresponding historical record of alarm events from the historical database and performing visual tracing display on the graphic control system interface.

[0008] As an optional implementation manner, it further includes:

[0009] Creating a domain division library during the configuration of the graphic control system, and assigning a domain identifier to each heterogeneous data source in the domain division library, and establishing a mapping relationship between the domain identifier and the corresponding sensor list, sampling frequency parameter, and production unit or factory area identifier;

[0010] In response to the reconstruction alarm signal within a certain domain exceeding a preset threshold, the reconstruction alarm signal and the historical reference data corresponding to the domain are sent to the cross-domain fusion classification module together to obtain a preliminary alarm classification result for the domain, and after obtaining the classification result, the domain identifier and the corresponding preliminary alarm classification result are written into the historical database.

[0011] As an alternative implementation, it further includes:

[0012] A distribution difference measurement unit is set in the cross-domain fusion classification module. The distribution difference measurement unit calculates the distribution difference degree based on the characteristics of the reconstruction alarm signals collected from each source domain and the target domain in the domain partition library and a preset divergence calculation method.

[0013] A discount coefficient calculation unit is also set in the cross-domain fusion classification module, which is used to generate a reliability or a discount coefficient according to the distribution difference degree.

[0014] After calculating the distribution difference degree and the discount coefficient, the alarm determination results of each source domain are weighted and fused, and the comprehensive alarm category and alarm level are output. Through a preset synthesis formula, a comprehensive determination result is obtained. When the determination result is higher than the preset fusion threshold, the comprehensive alarm category and alarm level are written into the historical database.

[0015] As an alternative implementation, it further includes:

[0016] A trust classification model is configured in the cross-domain fusion classification module. The trust classification model includes a nearest neighbor search unit and a composite class determination unit. The nearest neighbor search unit is used to identify the concentration of fault labels based on the neighborhood distribution of the reconstruction alarm signal in the high-dimensional feature space.

[0017] In response to the comprehensive alarm category obtained by the preset synthesis formula including multiple fault labels, the composite class determination unit is triggered to mark the alarm event as a composite alarm, and the source domain information, alarm category, alarm area, and timestamp of the composite alarm are recorded in the historical database, so that the graphic control system interface can trace and display the composite alarm event information when receiving a user query instruction.

[0018] As an alternative implementation, it further includes:

[0019] A PLC attribute domain evidence source is set in the cross-domain fusion classification module, and a health record table is stored for each PLC in the historical database.

[0020] Among them, the health record table includes the average scan cycle, communication timeout count, and hardware self-check result, which are used to indicate the operating state of the PLC.

[0021] Before calculating the distribution difference degree of the reconstructed alarm signal and fusing the synthesis formula, query the health record table of the PLC attribute domain evidence source and generate the corresponding attribute domain confidence;

[0022] In response to the health record table indicating that the PLC operating status is normal, increase the support degree for this alarm category;

[0023] In response to the health record table indicating that the PLC communication timeout or hardware failure is detected, reduce the support degree or confidence of the alarm category, and write the fusion result corrected by the PLC attribute domain evidence source into the historical database.

[0024] As an alternative implementation, it further includes:

[0025] Set the target PLC in the sparse reconstruction algorithm as the PLC that generates the compressed alarm signal and is triggered to the alarm state by the graphic control system;

[0026] Set up an association relationship table for the signal correlation degree between each PLC in the historical database to store the association parameters obtained based on similarity or mutual information;

[0027] Configure a multi-source dictionary management module in the sparse reconstruction algorithm for loading the historical waveform basis functions of the target PLC and the associated PLCs;

[0028] In response to the health value of the target PLC being less than the preset threshold, the multi-source dictionary management module references the corresponding mode from the waveform basis functions of the associated PLCs and reduces the weight of the local basis function of the target PLC in the reconstruction calculation to generate the reconstructed alarm signal;

[0029] In response to the health value of the target PLC being greater than or equal to the preset threshold, the multi-source dictionary management module performs the reconstruction calculation using the local basis function of the target PLC and outputs the reconstructed alarm signal;

[0030] After completing the generation of the reconstructed alarm signal, write the dictionary call information of the multi-source dictionary management module and the health records of the target PLC and the associated PLCs into the historical database.

[0031] As an alternative implementation, it further includes:

[0032] Install a vibration trigger for the monitoring object where the target PLC is located, and establish a vibration reference table for each vibration mode in the historical database to store the preset excitation frequency and the theoretical response curve;

[0033] When the graph control system is in a non-critical operation period, send an active vibration instruction to the vibration trigger and obtain the sampling data of the target PLC during the vibration period;

[0034] After comparing the sampling data with the theoretical response curve, if the deviation between the two is greater than the threshold, mark an abnormal active vibration detection in the health record table corresponding to the target PLC.

[0035] As an optional implementation manner, it further includes:

[0036] Set a vibration basis function library in the sparse reconstruction algorithm for storing the standard waveform basis functions corresponding to each excitation mode in the vibration reference table;

[0037] In response to the abnormal active vibration detection mark of the target PLC being written, apply a discount coefficient to the local basis function used by the target PLC in the reconstruction calculation, and increase the priority of the associated PLC basis function in the multi-source dictionary management module;

[0038] After the reconstruction is completed, write the reference information of the vibration basis function library and the abnormal active vibration detection mark of the target PLC into the historical database together.

[0039] As an optional implementation manner, it further includes:

[0040] Set a vibration excitation period configuration item in the database or configuration file for periodically triggering the vibration trigger and collecting the vibration response data of the target PLC;

[0041] In response to detecting that the amplitude or frequency of the vibration response of the target PLC deviates from the theoretical response curve more than the preset range multiple times within the excitation period, automatically reduce the health value of the target PLC to below a predetermined threshold;

[0042] After the health value is reduced, when performing cross-domain reconstruction on the compressed alarm signal uploaded by the target PLC based on the sparse reconstruction algorithm, call the PLC basis functions of other health values greater than or equal to the predetermined threshold to compensate for the data distortion of the target PLC, and write the final reconstruction result into the historical database.

[0043] Compared with the prior art, on the basis of ensuring the fidelity of the alarm signal, the present application uses undersampling and sparse reconstruction to significantly reduce the volume of alarm data. At the same time, combined with the preset area division rules and alarm type classification rules, it realizes mapping the reconstructed alarm signal to different alarm areas and alarm levels quickly and accurately. More importantly, while recording the alarm event level, area information, timestamp and processing status, the present invention forms a complete alarm history database through the association relationship with the PLC data points, and supports visual backtracking for different query requirements. It not only improves the processing efficiency of multi-source heterogeneous alarm signals in complex industrial sites, but also provides a reliable data basis for subsequent industrial safety analysis and equipment maintenance optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 is a schematic flowchart of the industrial safety alarm event recording and tracing method based on the graphic control system provided by the embodiment of the present application;

[0045] Figure 2 is a schematic diagram of the naming specification of various detectors provided by the embodiment of the present application;

[0046] Figure 3 is a schematic diagram of an example of the naming and annotation of detector-related points provided by the embodiment of the present application;

[0047] Figure 4 is a schematic diagram of the maintenance process of the health record form referring to the vibration unit provided by the embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0048] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments.

[0049] Please refer to Figure 1 , Figure 1 which is a schematic flowchart of the industrial safety alarm event recording and tracing method based on the graphic control system provided by the embodiment of the present application, including steps S101 to S104, where:

[0050] S101: Collect the original alarm signal, perform undersampling processing on the original alarm signal to generate a compressed alarm signal; for the compressed alarm signal, perform reconstruction through a sparse reconstruction algorithm to generate a reconstructed alarm signal;

[0051] S102: According to the preset area division rules and alarm type classification rules, divide the reconstructed alarm signal into alarm events of different alarm levels and different alarm areas;

[0052] S103: According to the divided alarm events, the alarm level, area information, occurrence time and processing status of the alarm event are automatically recorded in the historical database through the pre-associated PLC data points to form an alarm event history record;

[0053] S104: receiving a query instruction from a user, extracting corresponding alarm event history records from the history database according to the query instruction, and performing visual tracing display on the interface of the graphic control system.

[0054] When deploying the safety alarm event recording and tracing method based on the graphic control system described in the present invention at an industrial site, it is usually necessary to comprehensively consider multiple factors such as hardware environment, network communication, data storage, algorithm implementation, and human-computer interaction. The following takes the implementation of over-temperature alarm monitoring in a petrochemical production device as an example to illustrate the specific engineering implementation steps of the present invention.

[0055] First, at the hardware level, key equipment in petrochemical plants (such as reactors, separation towers, pipeline valves, etc.) are usually equipped with several temperature, pressure, flow and other sensors, and the data collected by the sensors are preliminarily processed or the switch quantity is judged through industrial-grade PLCs (such as Siemens S7 series, Schneider Modicon series or Rockwell Allen-Bradley, etc.). Real-time data exchange can be carried out between PLC and the host system through industrial Ethernet protocols (such as PROFINET, EtherNet / IP, Modbus TCP or OPC UA, etc.). In order to reduce network bandwidth pressure, reduce storage volume, and ensure timely capture of alarm information under critical working conditions, the present invention deploys undersampling and sparse reconstruction modules at the data acquisition end or in the middle edge computing node.

[0056] Specifically, a dedicated signal processing program is run on the field data acquisition station (an industrial computer or edge gateway device can be used), which regularly obtains the original alarm signal through the communication interface of the PLC. When the system detects that the temperature value of a certain measuring point exceeds the alarm threshold, it is considered to be in an "alarm state" and performs undersampling according to the sampling period set according to process needs. For example, for the overheating reaction stage with drastic temperature changes, the sampling period can be dynamically adjusted to 50ms; for the stable operation stage, it can be relaxed to 100ms or longer to reduce the amount of data without affecting the capture of key features. The undersampling process is usually combined with digital filtering and denoising to avoid unnecessary negative impacts of interference peaks on the reconstruction algorithm.

[0057] After obtaining the undersampled data, the signal processing program will call the sparse reconstruction algorithm library to complete the reconstruction of the alarm signal. This algorithm library can be implemented based on C / C++, Python or other high-performance languages to ensure real-time performance in the industrial field environment. The algorithms can include Orthogonal Matching Pursuit (OMP), K-SVD based on dictionary learning, and other reconstruction methods based on the theory of compressive sensing.

[0058] To improve the reconstruction effect, it is necessary to establish a dictionary library that conforms to the characteristics of the production process during system configuration. The dictionary library contains common pattern basis functions for temperature changes (such as slow rise, rapid jump, oscillation, etc.), and can also be continuously optimized through online learning during the initial operation stage. After being processed by the above sparse reconstruction algorithm, the reconstructed alarm signal obtained is relatively close to the original signal in terms of time resolution and amplitude accuracy, thus providing relatively comprehensive data support for subsequent alarm classification and trend analysis.

[0059] After the alarm signal reconstruction is completed, the system will determine the alarm event information according to the pre-configured area division and alarm type classification rules. Taking a petrochemical plant as an example, the whole plant may be divided into raw material storage area, reaction area, heat exchange area, finished product tank area, etc. according to geographical location or process flow; at the same time, different level threshold values can be set from level I to level III in the dimension of temperature alarm to distinguish early warning, fault and emergency states. When the reconstructed alarm signal corresponding to a certain sensor exceeds the level II threshold, the system will automatically generate a level II alarm event and identify the corresponding area. When implementing this process, it is usually necessary to maintain a mapping relationship table between "measurement point - area - alarm threshold - alarm level" in the system database or configuration file; once an alarm overrun is detected, the system can query this relationship table to quickly determine the area and level of the alarm. To ensure the flexibility of project implementation, operators can also dynamically modify this mapping relationship on the background interface of the graphic control system to adapt to the adjustment requirements of alarm thresholds for different production stages or seasonal working conditions.

[0060] After the alarm event is determined, the system will automatically write the core data of the alarm event (including alarm area, alarm level, generation time, alarm type, and the associated PLC data point ID, etc.) into the historical database. The historical database here can be either a traditional relational database (such as Oracle, SQL Server, MySQL) or a time series database (such as InfluxDB, TimescaleDB) to better meet the requirements of high-frequency writing and time series analysis.

[0061] In specific engineering implementations, the high availability and scalability of the database also need to be considered. For example, for large petrochemical integrated plants, it is recommended to deploy a dual-machine or cluster database solution to prevent the loss of production data caused by a single-point failure. At the same time, while the system writes alarm records, it can record user interaction information related to alarm handling, such as which operator confirmed or processed the alarm at what time, as well as operation instructions and interlock measures implemented during the handling process.

[0062] In the upper-level interface of a graphic control system (such as a SCADA or DCS system), the visualization and traceability function of the present invention is mainly realized by retrieving and presenting alarm events stored in the historical database. For a petrochemical site, the P&ID (Process Flow Diagram) or 3D digital twin model of the plant area can be associated with the alarm data. When the user clicks on a certain area or device, the system will display a list of the most recent alarm records of that device on the interface. The list content includes the alarm time, alarm level, handling status, and whether linkage measures have been taken, etc. If the user needs more detailed traceability analysis, they can click on the alarm record to enter the trend viewing mode and retrieve the temperature curve (or other relevant parameters such as pressure, flow rate, etc.) of the alarm in the minutes to hours before and after the trigger, so as to evaluate the cause and effect of the fault and ultimately provide reliable clues for root cause analysis for decision-makers.

[0063] In addition, in some application scenarios, operators may also need to perform statistical analysis and report output on alarm events, such as counting the total number of alarms in each area, the proportion of alarms at each level, as well as the average handling duration and response duration within a certain period of time. When implementing the present invention in engineering, a statistical function module can be added at the database level or the graphic control system level. Through aggregation queries or big data analysis techniques, a large number of historical alarms can be quickly summarized, and then visualized reports can be generated in the form of bar charts, pie charts, or line trend charts, or emails or text messages can be sent to management regularly. These functions can help plant managers understand the equipment health and process operation status, and provide a scientific basis for subsequent production improvement, equipment pre-maintenance, and personnel scheduling.

[0064] In this way, the present application effectively realizes the efficient compression, precise reconstruction, automatic event division, historical record management, and visualization and traceability of industrial field safety alarm signals. Among them, the undersampling and sparse reconstruction technology not only reduces the consumption of transmission and storage resources, but also ensures the retention and restoration of key features of the alarm waveform through appropriate dictionary design and online learning mechanisms. The function of automatically recording alarm events into the historical database and visual query improves the response efficiency and traceability analysis ability of operation and maintenance personnel for alarm events.

[0065] Exemplarily, the alarm group division rule can be:

[0066] Secondary group division rule: If there is only one, it is the GMS secondary group division rule. Check how many types of devices are in the menu, and divide into that many secondary alarm groups. For example, the detector corresponds to the alarm group GMS_DT, and the alarm light corresponds to the alarm group GMS_LAU, etc.;

[0067] Tertiary group division rule: For each secondary group, there will be alarm groups for several related buildings. For example, for the alarm group GMS_DT of the detector, there will be the following three tertiary alarm groups: GMS_FAB1_DT (all detectors in FAB1 will belong to this group), GMS_FAB2_DT (indicating that all detectors in FAB2 will belong to this group), GMS_SIH4YD_DT (indicating that all detectors in the silane station will belong to this group);

[0068] Quaternary group division rule: For each tertiary group, if it needs to be divided into smaller areas, it can be divided again. For example, the tertiary group GMS_FAB1_DT represents all detectors in FAB1. If FAB1 has two floors, it can be divided again into GMS_FAB1_F1_DT (all detectors on the first floor of FAB1 will belong to this group) and GMS_FAB1_F2_DT (all detectors on the second floor of FAB1 will belong to this group);

[0069] Quinary group division rule: For each quaternary group, if it needs to be divided into smaller areas, it can be divided again. For example, the quaternary group GMS_FAB1_F1_DT represents all detectors on the first floor of FAB1. If there are many machines on the first floor of FAB1 and there are many detectors in (or near) each machine, then it can be divided again as needed. For example, GMS_FAB1_F1_CVD1_DT (all detectors related to CVD1 on the first floor of FAB1 will belong to this group).

[0070] Exemplarily, for the naming rule of the alarm group related to the detector: GMS_DT (secondary alarm group naming), GMS_Large Area_DT (tertiary alarm group naming, the large area is generally the building name), GMS_Large Area_Medium Area_DT (quaternary alarm group naming, the medium area is generally the floor number), GMS_Large Area_Medium Area_Small Area_DT (quinary alarm group naming, the small area is generally a small room), GMS_Large Area_Medium Area_Small Area_Smaller Area_DT (sextary alarm group naming).

[0071] Except that the secondary alarm group is GMS_DT, the specific alarm groups behind can be divided layer by layer according to the area. The number of areas in different factories is different, which is manifested as different depths of the specific alarm groups.

[0072] Among them, GMS: General / Gas Monitoring System, a general / gas monitoring system;

[0073] FAB: Fabrication Facility, a manufacturing workshop or plant;

[0074] DT: Detector, a detector;

[0075] LAU: Local Alarm Unit, a local alarm unit (such as an alarm light or an alarm);

[0076] CVD: Chemical Vapor Deposition, a chemical vapor deposition equipment;

[0077] SIH4YD: Silane Yard, a silane station or a silane storage site;

[0078] F1, F2: Floor 1, Floor 2, representing the first floor and the second floor.

[0079] Exemplarily, please refer to Figure 2 , Figure 2 which is a schematic diagram of a naming specification for various detectors provided by an embodiment of the present application.

[0080] Exemplarily, please refer to Figure 3 , Figure 3 which is a schematic diagram of an example of naming and annotation for detector-related points provided by an embodiment of the present application.

[0081] As an alternative implementation, although the industrial safety alarm event recording and tracing method based on a graphic control system disclosed in the present invention can compress and sparsely reconstruct the original alarm signal, while reducing the burden of data transmission and storage, it retains the main feature information of the alarm signal. However, in some complex industrial scenarios, especially when the production process involves multiple types of devices, cross-regional or cross-plant linkages, the alarm signals often come from multiple heterogeneous data sources, and these data sources may be distributed in different production units or different plants (i.e., "multi-source domains"). Due to differences in process flow, operation methods, or environmental conditions, the distributions of data from different sources are inconsistent, which poses certain challenges to the accurate determination and grading of alarm events.

[0082] To further improve the accuracy of alarm event recognition and classification in a multi-source heterogeneous data scenario, the present invention can perform cross-domain auxiliary classification and fusion decision-making based on the alarm signal obtained by sparse reconstruction.

[0083] In specific implementation, in a typical industrial production environment, different factory areas or different production lines often deploy their own sensors and PLC systems. The alarm data output by these systems may vary in signal sampling frequency, data format, and measurement point definition, belonging to typical heterogeneous data sources.

[0084] For the convenience of subsequent cross-domain fusion processing, the present invention first conducts "domain" division on these different data sources during system configuration. Each domain can correspond to a factory area, a production line, or a subsystem (such as a heating furnace domain, a reaction area, a storage tank domain, etc.), and preliminary undersampling and sparse reconstruction operations are completed locally or at the edge.

[0085] When the system detects a potential alarm situation within a certain domain (such as parameters such as temperature, pressure, and flow exceeding the limit), the reconstructed alarm signal of this domain and historical reference data (labeled fault samples, normal samples, etc.) will be input into the cross-domain fusion classification module together. The purpose of this is to improve the judgment ability of the current alarm signal by leveraging the fault knowledge and classification models already accumulated in other domains (source domains) when there is a lack of sufficient training data or labeled information in the current target domain (Target Domain).

[0086] Furthermore, to better utilize the data of multiple source domains to assist in the alarm classification of the target domain, the present invention takes the following steps: compare the feature distribution of each source domain (such as the time-domain features, frequency-domain features, statistical features, etc. of the alarm signal) with the target domain; use methods such as K-L divergence (KL-Divergence), MMD (Maximum Mean Discrepancy), or JS divergence to measure the distribution difference; set a "reliability" or "discount factor" according to the size of the difference. If a source domain has a higher distribution similarity with the target domain, its reliability is higher, and vice versa, the discount factor is larger. In this way, in the scenario of joint assistance by multi-source domain data, the contribution of each source domain can be dynamically weighted, reducing the error impact caused by overly large distribution differences and improving the classification accuracy of cross-domain migration.

[0087] After obtaining the judgment results of each source domain regarding the "current target alarm signal", the present invention uses evidence theory fusion to comprehensively integrate the results of each source domain to obtain a more robust alarm classification output.

[0088] Specifically:

[0089] Each source domain classifies the target domain signal based on its existing fault / alarm model, such as judging whether it is a "temperature warning", "temperature fault", "valve switch abnormality", "comprehensive fault", etc.

[0090] For each source domain, in combination with the distribution difference between it and the target domain, calculate the confidence level (or "basic probability assignment") of the source domain;

[0091] Fuse the confidence levels of each source domain through the DS synthesis formula to obtain a comprehensive determination result. If the DS synthesis result is higher than the preset threshold, output the corresponding alarm category and level.

[0092] The advantage of using the DS rule for multi-source fusion is that it can better handle the possible conflicts and uncertainties between source domains. When the differences between some source domain data and the target domain are large, the weight of that source domain will be weakened through the aforementioned discount coefficient adjustment, thereby reducing the impact of misjudgment.

[0093] In addition, in some industrial scenarios, alarm types may often be coupled with each other (such as the coexistence of complex fault situations such as excessively high temperature and too low pressure). If only hard decisions are made (outputting a single alarm category at one time), it may sometimes bring a certain error rate. For this reason, the present invention further refines the identification of alarm categories after DS fusion, and the specific implementation can include the following:

[0094] Search for neighboring samples of the target alarm signal in the high-dimensional feature space. If the neighboring distribution is concentrated and there is a clear fault label, it is inclined to judge it as a single alarm category;

[0095] If the neighboring distribution is relatively dispersed or there is a coexistence of multiple fault labels, it may be judged as a composite alarm event (such as too high temperature + valve misoperation; simultaneous triggering of safety interlocks, etc.). The system will mark it as "composite type" in the alarm record and prompt the operator to conduct a comprehensive inspection of relevant equipment or operation links to reduce the possibility of missed alarms or false alarms.

[0096] Through this trust classification model, it is possible to more accurately distinguish single-class faults and composite faults on the basis of cross-domain migration classification results, and also take into account various complex alarm scenarios, improving the safety guarantee level of actual applications.

[0097] After completing the classification and level confirmation of alarm events, the present invention will uniformly write information including "cross-domain fusion classification output", "alarm category", "alarm area", "timestamp", "processing status", etc. into the historical database. For multi-source heterogeneous data, the determination processes of each source domain (such as source domain ID, confidence level, discount coefficient) and the final DS fusion result can also be additionally recorded for future retrospective analysis and algorithm performance evaluation.

[0098] When the user issues a query instruction on the graphic control system interface, the system can display alarm records including the following key information on the visualization interface:

[0099] The alarm signal waveform of the target domain (after sparse reconstruction), the determination results and their confidence levels of multiple source domains, the final classification result output by DS fusion, the determination process of this alarm event under the trust classification model (single class or composite class), and the processing status (unprocessed, processing, or processed).

[0100] If it is found in subsequent investigations that there are situations such as "cross-domain misjudgment" or "overestimation / underestimation of distribution differences" in some events, the correlation parameters between the source domain and the target domain can also be adjusted in the system background, and the correction results are written into the model configuration again, so as to continuously optimize the migration classification effect between domains.

[0101] Exemplarily, in actual deployment, the multi-source heterogeneous data fusion and migration classification module can be deployed on the edge side or in the cloud according to requirements. For scenarios with high real-time requirements, the corresponding algorithm modules and fusion strategies can be configured on the on-site industrial control computer so that classification and linkage can be completed immediately when an emergency alarm occurs; for decentralized production devices with a large number of alarms, cross-domain fusion and offline analysis can be completed on the cloud server or data center, which not only reduces the burden on edge nodes but also better summarizes the alarm information of multiple factory areas, realizing centralized monitoring and historical big data analysis over a larger range.

[0102] In occasions with high requirements for industrial safety such as petrochemical, iron and steel smelting, power systems, and pharmaceuticals, based on traditional sparse reconstruction, the present invention can not only make better use of the fault knowledge bases between different domains and reduce the decline in classification accuracy caused by insufficient training samples, but also effectively identify composite alarms through DS fusion and the trust classification model, reducing the false alarm and missed alarm rates in industrial sites and significantly improving the production safety level.

[0103] In this way, the present invention can more effectively identify alarm categories and make credible decisions in multi-source domain scenarios, reasonably discount the distribution differences between different source domains, and finally further refine alarm events into single classes or composite classes with the help of the trust classification model, enhancing the adaptability of the alarm management system to actual complex working conditions.

[0104] As an alternative implementation, in order to address the misjudgment problem caused by the acquisition error of PLCs in industrial sites due to network latency or hardware failures, an additional attribute domain for the operating status of PLCs is set on the basis of the original cross-domain fusion classification module and regarded as an independent evidence source.

[0105] Specifically, first in the system configuration phase, a status monitoring mechanism is established for each PLC within the domain, collecting information such as its real-time operating mode, hardware self-check results, and communication error rate, and maintaining a PLC health table in the database. Subsequently, when performing cross-domain alarm fusion, certain basic probability assignments are respectively given to the determination results of each conventional source domain to represent the support or confidence of the source domain for a certain alarm category; at the same time, the PLC attribute domain will also generate a set of basic probability assignments corresponding to the alarm category according to the current PLC health status. If it is detected that the PLC is operating normally and the collected network communication quality is within the set threshold range, the support for each alarm category will be increased when determining the target alarm signal; if it is found that the PLC has an abnormal scan cycle or the fault diagnosis indicates that the hardware is running unstably, the support for this alarm category will be reduced or directly regarded as conflicting evidence.

[0106] In this way, the original method of simply performing DS synthesis relying only on the fault or alarm models output by each source domain is improved to introduce the credibility correction of the PLC attribute domain before fusion, thereby avoiding the situation of still considering the collected data to have the same credibility as other normal domains when the PLC hardware or network is abnormal.

[0107] During the DS synthesis process, if the PLC attribute domain indicates normal, the recognition degree of the alarm determination result of the same domain can be significantly enhanced; conversely, if it is detected that the credibility output by the PLC attribute domain is low, it means that there is a large interference or distortion in the alarm signal obtained by this domain, and finally the DS synthesis formula will discount its evidence and give it a lower weight.

[0108] After the fusion is completed, the system will not only output the synthesized alarm category and level, but also store the health information and the corresponding conflict or support evidence used when the PLC attribute domain participates in the calculation in the historical database for subsequent backtracking and evaluation.

[0109] Through this method, it is possible to better eliminate or weaken the interference from abnormal PLCs in the cross-domain fusion determination of multiple source domains, reduce false alarms or missed alarms caused by hardware or network abnormalities in a single source domain, and thus make the alarm classification results for complex working conditions more reliable and stable.

[0110] Exemplarily, in the reaction zone of a large petrochemical plant, three industrial-grade PLCs are deployed, with models Siemens S7-400, Schneider Modicon M340, and Rockwell Allen-Bradley CompactLogix respectively. Each PLC is connected to the host computer through industrial Ethernet. To ensure accurate monitoring of key parameters such as the temperature and pressure of the reactor, each PLC adopts a scanning cycle of 0.1 second and regularly uploads data to the on-site edge computing gateway. A dedicated signal processing software and a cross-domain fusion classification module are installed on the edge gateway. This software is developed by mixing Python and C++ and can uniformly manage the alarm signals from different PLCs. When the system is deployed, an independent "PLC health status" record is established in the database for each PLC, including PLC_ID, average scanning cycle, communication timeout count, hardware self-check result, and the reserved fault log path. The health status is represented by a floating-point number from 0 to 1, and the closer it is to 1, the more stable the PLC is. The database is deployed in the local area network server in the factory in the form of a MySQL cluster, and real-time insertion and historical retrieval interfaces are provided at the same time.

[0111] During the on-site operation process, when one of the PLCs monitors that the temperature of the reactor exceeds the set level-II alarm threshold (such as 250 degrees Celsius), it will upload the data stream of the temperature measurement point together with basic fields such as the time stamp and over-limit information to the edge gateway. The signal processing software on the edge gateway first performs downsampling and sparse reconstruction on this set of temperature values to generate a reconstructed alarm signal, and then calls the cross-domain fusion classification module for preliminary analysis. The module will compare this alarm signal with the existing fault models in other domains (such as the heat exchange area, storage tank area), and obtain the basic probability assignment for the temperature alarm through the DS synthesis rule. During this process, the system additionally introduces the PLC attribute as an independent evidence source for weighting. The specific method is as follows:

[0112] After obtaining the alarm data transmitted by the PLC, the cross-domain fusion classification module reads the current health status value and interlock trigger status of this PLC from the database. If the health status value is greater than 0.8 and no interlock trigger occurs, a higher degree of support, such as between 0.6 and 0.7, will be assigned to the "high-temperature alarm" category of this PLC in this alarm; if it is detected that the communication timeout count of this PLC is high, the hardware self-check is incorrect, or the interlock trigger conflicts with the current alarm category, the degree of support for the corresponding alarm category will be reduced to below 0.2 or even marked as conflicting evidence. This degree of support value, together with the determination results of other source domains, generates the final confidence level of the "temperature fault alarm" through the Dempster synthesis formula.

[0113] Once the confidence level of the "temperature fault alarm" output by the cross-domain fusion classification module is higher than the pre-set threshold of 0.8, the system confirms that this alarm is a valid fault, automatically records fields such as "Alarm Category: Temperature Fault", "Alarm Level: Level II", "PLC Attribute Domain Judgment: High Health", and "Processing Status: Unprocessed" in the historical database, triggers a flashing prompt on the graphic control interface at the same time, and generates an alarm log for the operator to sign for processing.

[0114] In another alarm scenario, if the health of a certain PLC drops to 0.4 and the hardware diagnosis shows that the analog input module is abnormal, the system will, based on the evidence output by this attribute domain, identify this alarm signal as a less credible alarm source, and the DS synthesis formula will correspondingly reduce the final support for this alarm type. If the fusion result is lower than the threshold value, the system will only mark it as "pending review" in the alarm list and prompt the operator to give priority to repairing this PLC or confirming whether its communication status is normal. The judgment result of this attribute domain will also be stored in the database as an independent "fusion auxiliary record" so that when the operation and maintenance personnel query the running status of the PLC when a specific alarm occurs, they can see detailed health values, communication timeout counts, and final discount coefficients and other information, thus providing reliable data support for alarm analysis and fault troubleshooting.

[0115] In the above way, multiple on-site PLCs can stably provide reliable alarm data under normal circumstances. Once a hardware fault or network jitter occurs, the alarm evidence of the PLC will be automatically discounted during DS synthesis and will not cause significant interference to the overall alarm judgment. Using the PLC attribute as an independent evidence source can significantly reduce the risk of false alarms or missed alarms caused by a single-point PLC abnormality in the multi-source domain fusion scenario and ensure the safety interlock and process operation stability of the petrochemical plant.

[0116] As an optional implementation method, in some complex industrial scenarios, especially when the PLC sampling periods are not exactly the same or there are delays in the on-site network, the alarm signals uploaded from each source domain may have problems such as time series misalignment, missing point filling, or large differences in sampling rates, resulting in an inability to accurately reflect the true differences between the source domain and the target domain when directly comparing the feature distributions of different domains.

[0117] Therefore, a divergence calculation method that combines time series alignment and local weighting strategies is adopted. By performing dynamic registration of data before divergence calculation, it is possible to solve the problem of classification inaccuracy that may be caused by time series mismatch in the graphic control system.

[0118] In specific implementation, first, the reconstructed alarm signals of each source domain are subjected to timestamp standardization in the cross-domain fusion classification module, and the feature vectors are resampled or interpolated at an adjustable time step, so that the sampling time series of the source domain coincides as much as possible with that of the target domain on the same time axis. If abnormal reports or communication delay records are reported by the PLC at some sampling points, the data in this section are weighted interpolated or removed at this stage. Subsequently, a local weighted calculation method is introduced for the realigned feature vectors: a higher weight is given to the time period or the feature peak section closer to the alarm trigger point of the target domain; for the section far from the critical time period or with a large interpolation error, a decay coefficient is used to reduce its influence. By combining this local weighting with time alignment, while retaining the overall distribution comparison, the signal similarity between the source domain and the target domain at critical time points can be measured more accurately, so as to obtain a corrected "weighted divergence value".

[0119] At the implementation level, it can be modified based on the MMD (Maximum Mean Discrepancy) framework: when calculating the kernel function, not only the Euclidean distance or kernel mapping similarity of the feature points between the source domain and the target domain is considered, but also the temporal difference and local weight factor after time alignment are introduced.

[0120] Specifically, when calculating a pair of feature points, its timestamp is compared with that of the feature point of the target domain. If the time difference between the two is small and close to the alarm peak position, the output of its kernel function is increased; if the time difference between the two is large or there are many missing data segments in this source domain data section, the contribution of the kernel output of this point pair is reduced. After summing the kernel functions, the obtained MMD value can better reflect the real inter-domain difference. Finally, when the system performs cross-domain fusion classification, a new reliability or discount coefficient is generated based on this weighted divergence value. A higher confidence level is assigned to the source domain with a lower divergence value (indicating that the two domains are very similar at the critical period after alignment), and vice versa, the discount coefficient is increased to weaken its influence.

[0121] In specific implementation, it is necessary to add a "time-series registration" policy configuration item in the database or configuration file to record the sampling period and network delay baseline of each PLC, facilitating dynamic compensation of the feature sequence based on the real-time measured communication delay after alarm triggering. The time alignment offset and local weighted distribution during each divergence calculation process can also be saved in the historical database, enabling operation and maintenance personnel to clearly see information such as "the alignment error between this domain and the target domain within a certain time range" and "the interpolation and rejection processing carried out in which sampling segments" during retrospective analysis, so as to better understand the final formation process of the divergence value. If a certain PLC has serious data anomalies at a certain stage, the system can also discover a large number of high-interpolation or high-rejection records based on the time-series registration results, and then further reduce the support degree of this domain through the PLC attribute domain evidence source, avoiding misleading the final alarm fusion decision.

[0122] In this way, when the graph control system performs multi-source domain cross-domain fusion determination, it can not only quantify the simple distribution differences, but also make more delicate dynamic corrections for situations unique to the industrial site, such as inconsistent sampling periods, network jitters, and local PLC anomalies, reducing misjudgments caused by time-series misalignment or data loss. When cooperating with the aforementioned scheme for the PLC attribute domain, this improved divergence calculation can further enhance the identification and discounting of abnormal data, improving the stability and accuracy of alarm classification of the graph control system under complex working conditions.

[0123] As an alternative implementation method, it further includes:

[0124] Set the target PLC in the sparse reconstruction algorithm as the PLC that generates the compressed alarm signal and triggers the alarm state by the graph control system;

[0125] Set an association relationship table for the signal correlation degree between each PLC in the historical database to store the association parameters obtained based on similarity or mutual information;

[0126] Configure a multi-source dictionary management module in the sparse reconstruction algorithm for loading the historical waveform basis functions of the target PLC and associated PLCs;

[0127] In response to the health degree value of the target PLC being less than the preset threshold, the multi-source dictionary management module references the corresponding mode from the waveform basis functions of the associated PLCs and reduces the weight of the local basis function of the target PLC in the reconstruction calculation to generate the reconstructed alarm signal;

[0128] In response to the health degree value of the target PLC being greater than or equal to the preset threshold, the multi-source dictionary management module performs the reconstruction calculation using the local basis function of the target PLC and outputs the reconstructed alarm signal;

[0129] After the reconstruction alarm signal is generated, the dictionary call information of the multi-source dictionary management module and the health records of the target PLC and associated PLCs are written into the historical database together.

[0130] The compressive sensing and sparse reconstruction described above mainly rely on a single signal or the historical feature library in the same domain (such as each measuring point under the same PLC) to complete the reconstruction. However, in the actual industrial field graphic control system, if the operating state of a certain PLC is unstable or the measurement accuracy of some sensors decreases, when simply relying on its local historical data for reconstruction, the reconstruction result may be distorted or the recognition of sudden fault features may be insufficient. Therefore, on the basis of the original compression and sparse reconstruction, a dynamic multi-source dictionary or cross-PLC collaborative reconstruction strategy can be designed by combining the correlation relationship between multiple PLCs and the health information of each PLC, so as to maintain a high reconstruction quality and improve the stability of alarm recognition when the data credibility of a certain PLC decreases.

[0131] In specific implementation, in the system configuration stage, in addition to establishing a health record table for each PLC (including the average scan cycle, communication timeout count, hardware self-check result, fault log, etc.), a PLC association relationship table for the signal association degree existing between each PLC will also be established in the database. This association relationship can be calculated based on methods such as the time series similarity, covariance, or mutual information of historical data, and stored in the form of a matrix or list.

[0132] For example, in a petrochemical production device, multiple PLCs respectively monitor similar process units (such as adjacent reactors and heat exchangers). Since their temperature or pressure changes synchronously with the process, they often show a high correlation coefficient; if some PLCs are far apart geographically or in the process flow, their signal association degree may be low.

[0133] When a certain PLC detects an alarm state and uploads the compressed alarm signal, before the system executes the step of "reconstructing the compressed alarm signal through a sparse reconstruction algorithm", it will additionally query the current health value of this PLC and its association relationship with other PLCs. If the health value is low (such as less than 0.5) and the communication timeout count increases, indicating that there may be a data reliability problem with this PLC, the system will increase the reference weight of the historical samples provided by other PLCs with a higher association degree and a normal health level (such as greater than 0.8).

[0134] The specific approach is that when building the dictionary library required for sparse reconstruction, in addition to loading the historical mode basis functions of this PLC, a part of the typical fault / normal sample waveforms from other PLCs with a relatively high correlation will be selected to expand the dictionary or increase the priority of the corresponding basis functions. In this way, if the local data of this PLC is incomplete due to faults or interference, the reconstruction algorithm can also obtain useful patterns from the data of other PLCs with higher health and higher similarity, making the reconstructed alarm signal closer to the real situation in terms of amplitude and timing.

[0135] At the software implementation level, a dynamic dictionary management module can be maintained on the edge computing node or industrial control server. This module first reads the PLC correlation table and the health information of the target PLC during reconstruction. If the health of the target PLC is higher than a certain threshold (such as 0.8), the historical waveform basis functions of this PLC itself are mainly used; if the health is between 0.5 and 0.8, the waveform basis functions of the associated PLCs will be moderately introduced to reduce the noise or distortion that may be brought by the local data; if the health is lower than 0.5 and communication failures occur frequently, the waveform basis functions from PLCs with high health will be used as the main reference, and a discount factor will be applied to the basis functions of this PLC (reducing the weight during the dictionary search process), so as to ensure the accuracy of the reconstruction result to the greatest extent. This process can be achieved through weighted sparse coding or multi-source dictionary fusion. For example, during the K-SVD dictionary update or orthogonal matching pursuit (OMP) process, higher priority is assigned to the basis functions of the associated PLCs.

[0136] At the hardware deployment level, sufficient computing and storage resources can be configured on the edge computing node (such as an industrial PC or a dedicated AI gateway) for online execution of multi-source dictionary learning and sparse reconstruction. If the system scale is large and the data of multiple PLCs needs to be processed simultaneously, distributed dictionary management and reconstruction algorithms can be adopted on the server or in the cloud, and the real-time status (including PLC health, correlation, etc.) is read from the database cache (such as Redis or memory table) to reduce the latency of querying the relational database. For scenarios with extremely high real-time requirements on site, a minimized dictionary can be retained locally for fast matching, while complex multi-source updates or global dictionary training are performed on the backend server periodically.

[0137] In this way, when a certain PLC is abnormal, the alarm signal compression and reconstruction process no longer relies solely on the historical basis functions of this PLC, but uses the healthy data of the highly associated PLCs to assist in correction, achieving a more accurate restoration of the faulty or distorted segments. Compared with the traditional single-PLC dictionary reconstruction, this solution can better alleviate the reconstruction deviation caused by single-point faults, and in occasions such as petrochemical or power systems where multiple PLCs are interconnected, it can improve the overall alarm reconstruction accuracy, making the alarm determination obtained in the cross-domain fusion classification stage more reliable and more in line with the actual process requirements.

[0138] Exemplarily, when the temperature channel of a certain PLC drifts for a long time or the noise soars due to a hardware failure, other PLCs or sensors with a high degree of association with the PLC can still supplement the reference waveforms in a similar process state for the target PLC through the data they normally collect, making the sparse reconstruction of the compressed signal capture peaks and trends more accurately; conversely, when the system monitors that the health of the associated PLC itself is also declining, the dependence on its data will be reduced to avoid mistakenly using abnormal values as a benchmark for reconstruction. Thus, in the mode of multiple redundancies and cross-PLC collaboration in the industrial field, it is possible to obtain a relatively high-quality reconstructed result of the alarm signal even in the presence of differences in equipment health status and multi-source heterogeneous signals, solving the problem that traditional single-source sparse reconstruction is prone to distortion or data loss when the PLC is abnormal.

[0139] Please refer to Figure 4 , Figure 4 which is a schematic diagram of a maintenance process for a health record form that references a vibration unit provided in an embodiment of the present application, including steps S201 to S203, where:

[0140] S201: Install a vibration trigger for the monitoring object where the target PLC is located, and establish a vibration reference table in the historical database for each vibration mode to store the preset excitation frequency and theoretical response curve;

[0141] S202: When the graphic control system is in a non-critical operation period, send an active vibration instruction to the vibration trigger and obtain the sampling data of the target PLC during the vibration period;

[0142] S203: After comparing the sampling data with the theoretical response curve, if the deviation between the two is greater than the threshold, mark an abnormal active vibration detection in the health record form corresponding to the target PLC.

[0143] As an optional implementation manner, a vibration basis function library can be set in the sparse reconstruction algorithm to store the standard waveform basis functions corresponding to each excitation mode in the vibration reference table;

[0144] In response to the writing of the abnormal active vibration detection mark of the target PLC, apply a discount coefficient to the local basis function used by the target PLC in the reconstruction calculation, and increase the priority of the basis function of the associated PLC in the multi-source dictionary management module;

[0145] After the reconstruction is completed, write the reference information of the vibration basis function library and the abnormal active vibration detection mark of the target PLC into the historical database together.

[0146] As an alternative implementation, a vibration excitation period configuration item can be set in the database or configuration file to periodically trigger the vibration trigger and collect vibration response data of the target PLC;

[0147] In response to detecting that the amplitude or frequency of the vibration response of the target PLC deviates from the theoretical response curve by more than a preset range multiple times within the excitation period, automatically reduce the health value of the target PLC to below a predetermined threshold;

[0148] After the health value is reduced, when performing cross-domain reconstruction on the compressed alarm signal uploaded by the target PLC based on the sparse reconstruction algorithm, call the basis functions of other PLCs whose health values are greater than or equal to the predetermined threshold to compensate for the data distortion of the target PLC, and write the final reconstruction result into the historical database.

[0149] Based on existing solutions such as multi-source dictionaries and cross-PLC collaborative reconstruction, this application further utilizes an active vibration trigger and signal feedback acquisition mechanism to provide an additional reference signal for the PLC, aiming to improve the reconstructed alarm signal generated by reconstructing the compressed alarm signal through the sparse reconstruction algorithm, thereby enhancing the accuracy and calibratability of the data collected by the sensor or PLC under certain special working conditions.

[0150] In the above-mentioned solution, the PLC only triggers alarms and reports data based on the process variables monitored by the sensor (such as temperature, pressure, liquid level, flow rate, etc.). When potential failures occur in the sensor or PLC (such as decreased sensitivity, reference drift, range damage) but are not detected in time, the uploaded data will affect the subsequent compressive sampling and sparse reconstruction quality, which may lead to alarm errors.

[0151] Therefore, this application deploys a controllable vibration unit near the PLC (or on the mechanical structure of the monitored object). This vibration unit is not an essential part of the production process but is specifically used to actively generate vibration excitation at specific moments and let the sensor capture the corresponding feedback signal of this vibration. By comparing the measured vibration signal with the expected theoretical waveform or historical reference, the health or deviation of the sensor and its corresponding PLC channel can be evaluated; once it is found that the characteristic response of the sensor differs significantly from the standard response, the data weight of this PLC can be discounted or corrected accordingly during the subsequent reconstruction process, or the matching process of the dictionary library can be calibrated.

[0152] At the data level, the system needs to add a "Active Vibration Reference Table" for each PLC in the database. This table contains vibration mode ID, excitation frequency, excitation amplitude, corresponding theoretical response curve, and applicable measurement point range. According to the rigidity, material properties, or structural form of different on-site equipment, multiple sets of vibration excitation schemes can be configured. For example, the vibration frequencies of thicker and thinner pipes are different, or the induction frequency bands of temperature sensors and pressure transmitters also vary. Once a certain vibration mode ID is selected, the PLC will trigger the vibration unit to perform excitation at a specified moment according to the corresponding instructions, and collect the entire vibration response data from the sensor end.

[0153] At the hardware level, a vibration trigger and a drive unit can be installed near the on-site equipment or the PLC cabinet. This vibration trigger can use piezoelectric ceramic actuators, electromagnetic vibrators, or other industrial vibration devices, and can generate mechanical vibrations with a certain frequency, amplitude, or waveform after the instruction is issued. When the PLC executes the "Active Vibration Test" instruction, it will first record the vibration start time, and then collect sensor signals at high frequency within a preset time window (such as within a few seconds), and preliminarily compare this vibration response with the reference curve. If the comparison result shows that the deviation value exceeds a certain threshold, such as "the vibration peak drop is greater than 20%" or "the main resonance frequency offset exceeds a certain standard", it indicates that there may be attenuation or calibration inaccuracy in the sensor or the PLC input module. At this time, the PLC will write an "Active Vibration Detection Abnormality" flag into the health table, and the edge computing or the server will appropriately reduce the weight of the data provided by this PLC during subsequent compression and reconstruction.

[0154] At the software implementation level, in addition to the above health flags, the vibration response record can also be incorporated into the dictionary library for sparse reconstruction.

[0155] The specific approach is as follows: When configuring the system, establish a corresponding "Standard Waveform Basis Function" for each vibration mode ID, and after the PLC completes the vibration test, write the actually collected "Measured Waveform" into the "Vibration Comparison Record Table" as well. If the actual measurement and the standard are not very different, it can be considered that the current state of this measurement point is good; if the difference is significant, update the historical basis function under this PLC or mark it as "low confidence" to avoid subsequent reconstruction algorithms relying heavily on this inaccurate basis function. Further optimization is that this "active vibration" can be initiated periodically under normal working conditions to perform online detection of the drift of the sensor and the PLC input channel, forming a set of "dynamic calibration" processes; once it is detected that the inaccuracy accumulates to a certain extent, the operation and maintenance personnel can decide whether to perform on-site maintenance or sensor replacement.

[0156] At the algorithm processing level, combined with the above-mentioned solutions for cross-PLC association and health degree discount, when the target PLC is marked with a low health degree due to abnormal vibration test results, the system can further lower the weight of the local historical waveform basis function extracted from this PLC during reconstruction. At the same time, if other high-health-degree PLCs have performed well historically under the same or similar vibration modes, more reference modes can be supplemented from the basis functions of these PLCs to ensure accurate signal reconstruction when an alarm state occurs.

[0157] For the usage scenario of multi-source dictionaries, if the vibration mode IDs are the same and the device structures are similar, the vibration waveform features can be shared among multiple PLCs, so that when a certain PLC performs poorly, the vibration features of other healthy PLCs can be used to make up for it. The reconstructed alarm signal obtained in this way is closer to the true fault state in terms of features such as amplitude and phase, reducing the distortion phenomenon caused by sensor drift or hardware failure.

[0158] Exemplarily, an electromagnetic vibrator is installed on a certain pipeline section. The system issues a trigger command during a non-critical operation period, causing the vibrator to vibrate at a frequency of 150 Hz for 2 seconds. The PLC captures the response curve through a temperature sensor or an acceleration sensor and stores it in the "Vibration Comparison Record Table". After comparing this curve with the standard curve, if the peak deviation is small and the main harmonics are consistent, the health degree of the PLC is not affected; if abnormalities such as a peak deviation > 30% and the main harmonics being misaligned by more than 0.5 times occur, the health degree of this PLC is lowered to 0.6 and the database is updated. Subsequently, if this PLC issues a real fault alarm on the same day, the compressed alarm signal it uploads will process the local basis function in a discounted manner during sparse reconstruction, and more basis functions of other associated PLCs will be referenced, avoiding using sensor data with serious drift as a reliable basis, thereby reducing the probability of false alarms or missed alarms.

[0159] In this way, the "comparison waveform" collected by the active vibration trigger can directly or indirectly participate in the optimization of PLC health degree evaluation and sparse reconstruction algorithms, fundamentally reducing the alarm data quality problems caused by sensor inaccuracy or hardware failures. For large equipment groups or remote distributed PLCs, it can significantly improve the efficiency of fault prediction and online diagnosis, enabling the graphic control system to still maintain high-quality reconstruction and accurate classification of alarm signals under complex working conditions. Especially in occasions that require high-precision monitoring of vibration, noise or other dynamic characteristics (such as rotating machinery, vibrating screening devices, etc.), this solution can effectively solve the problem that it is difficult to detect the attenuation of sensor accuracy in a passive observation mode, allowing the operation and maintenance personnel to more actively master the actual state and real-time correct the basis function library of compressive sensing and reconstruction algorithms.

[0160] It should be noted that during the application of the solution of the present invention, the required software and hardware configurations can all be implemented based on the common PLC control system and data acquisition devices in industrial sites, and can be flexibly expanded in combination with existing edge computing or cloud server deployment methods. Whether it is a small-scale demonstration device within a single factory area or a large-scale distributed industrial system across regions and factory areas, various process flows and actual production requirements can be adapted by flexibly configuring functional modules such as "vibration trigger strategy", "multi-source dictionary management module", and "PLC health record form". By seamlessly integrating with the on-site process control logic, this solution can provide more detailed alarm data basis and accurate fault identification results for operators or system maintenance personnel without significantly increasing the production operation burden, thus achieving an effective balance between stable and safe operation and rapid fault diagnosis.

[0161] The solution described in the present invention effectively combines various methods such as compressive sensing and sparse reconstruction, DS fusion inference, multi-source PLC association, and active vibration calibration at the technical level, and strings together the acquisition, reconstruction, fusion, and health assessment of alarm data into a systematic process. Through the optimized design for industrial sensor characteristics in the vibration trigger and health management mechanism, the present invention not only reduces the probability of false alarms or missed alarms caused by sensor failures, PLC anomalies, or network delays, but also makes the results of compressive reconstruction closer to the dynamic changes of the actual production process. Coupled with the coordinated use of technical means such as time series registration and local weighted divergence in the multi-source fusion stage, the entire alarm management process has higher robustness and scalability, and can be widely applied to occasions with high requirements for alarm accuracy and real-time performance.

[0162] It should be understood that in various embodiments of the present application, the magnitudes of the serial numbers of the above processes do not mean the order of execution. The order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present application.

[0163] The preferred embodiments of the present invention disclosed above are only used to help explain the present application. The preferred embodiments do not describe all the details in detail, nor do they limit the specific implementation manner of the application. Obviously, many modifications and changes can be made according to the content of this specification. These embodiments are selected and specifically described in this specification to better explain the principle and practical application of the present application, so that those skilled in the art can understand and utilize the present application well. The present application is only limited by the claims and their full scope and equivalents.

Claims

1. An industrial safety alarm event recording and tracing method based on a graphic control system, characterized in that, Including: Collecting the original alarm signal and performing undersampling processing on the original alarm signal to generate a compressed alarm signal; Reconstructing the compressed alarm signal through a sparse reconstruction algorithm to generate a reconstructed alarm signal; According to the preset area division rule and alarm type classification rule, dividing the reconstructed alarm signal into alarm events with different alarm levels and different alarm areas; According to the divided alarm events, automatically recording the alarm level, area information, occurrence time, and processing status of the alarm events into the historical database through pre-associated PLC data points to form an alarm event historical record; Receiving a query instruction from the user, extracting the corresponding alarm event historical record from the historical database according to the query instruction, and performing visual trace display on the graphic control system interface; The method further includes: Creating a domain division library during the configuration of the graphic control system, and assigning a domain identifier to each heterogeneous data source in the domain division library, and establishing a mapping relationship between the domain identifier and the corresponding sensor list, sampling frequency parameter, and production unit or plant identifier; In response to the reconstructed alarm signal in a certain domain exceeding a preset threshold, sending the reconstructed alarm signal and the historical reference data corresponding to the domain to the cross-domain fusion classification module to obtain a preliminary alarm classification result for the domain, and after obtaining the classification result, writing the domain identifier and the corresponding preliminary alarm classification result into the historical database; The method further includes: Setting a distribution difference measurement unit in the cross-domain fusion classification module, and the distribution difference measurement unit calculates the distribution difference degree based on a preset divergence calculation method according to the characteristics of the reconstructed alarm signals collected from each source domain and the target domain in the domain division library; A discount coefficient calculation unit is further provided in the cross-domain fusion classification module for generating a reliability or discount coefficient according to the distribution difference degree; After calculating the distribution difference degree and the discount coefficient, performing weighted fusion on the alarm determination results of each source domain, outputting a comprehensive alarm category and alarm level, and obtaining a comprehensive determination result through a preset synthesis formula. When the determination result is higher than a preset fusion threshold, writing the comprehensive alarm category and alarm level into the historical database.

2. The industrial safety alarm event recording and tracing method based on a graphic control system according to claim 1, characterized in that, It further includes: Configuring a trust classification model in the cross-domain fusion classification module, the trust classification model includes a nearest neighbor search unit and a composite class determination unit, and the nearest neighbor search unit is used to identify the concentration of fault labels based on the neighborhood distribution of the reconstructed alarm signal in the high-dimensional feature space; In response to the comprehensive alarm category obtained by the preset synthesis formula including multiple fault labels, triggering the composite class determination unit to label the alarm event as a composite alarm, and recording the source domain information, alarm category, alarm area, and timestamp of the composite alarm into the historical database, so that the graphic control system interface can trace and display the composite alarm event information when receiving a user query instruction.

3. The industrial safety alarm event recording and tracing method based on a graphic control system according to claim 2, wherein It further includes: Setting a PLC attribute domain evidence source in the cross-domain fusion classification module, and storing a health record table for each PLC in the historical database; Among them, the health record table includes the scanning period mean value, communication timeout count, and hardware self-check result, which are used to indicate the operating state of the PLC; Before calculating the distribution difference degree of the reconstructed alarm signal and fusing the synthesis formula, query the health record table of the PLC attribute domain evidence source and generate the corresponding attribute domain confidence; In response to the health record table indicating that the PLC operating state is normal, increase the support degree for this alarm category; In response to the health record table indicating that the PLC communication timeout or hardware failure is detected, reduce the support degree or confidence of the alarm category, and write the fusion result corrected by the PLC attribute domain evidence source into the historical database.

4. The industrial safety alarm event recording and tracing method based on the graphic control system according to claim 3, characterized in that, It also includes: Set the target PLC in the sparse reconstruction algorithm as the PLC that generates the compressed alarm signal and is triggered to the alarm state by the graphic control system; Set an association relationship table for the signal association degree between each PLC in the historical database to store the association parameters obtained based on similarity or mutual information; Configure a multi-source dictionary management module in the sparse reconstruction algorithm, which is used to load the historical waveform basis functions of the target PLC and the associated PLCs; In response to the health degree value of the target PLC being less than the preset threshold, the multi-source dictionary management module references the corresponding mode from the waveform basis functions of the associated PLCs, and reduces the weight of the local basis function of the target PLC in the reconstruction calculation to generate the reconstructed alarm signal; In response to the health degree value of the target PLC being greater than or equal to the preset threshold, the multi-source dictionary management module performs the reconstruction calculation using the local basis function of the target PLC and outputs the reconstructed alarm signal; After completing the generation of the reconstructed alarm signal, write the dictionary call information of the multi-source dictionary management module and the health records of the target PLC and the associated PLCs into the historical database.

5. The industrial safety alarm event recording and tracing method based on the graphic control system according to claim 4, characterized in that It also includes: Install a vibration trigger for the monitoring object where the target PLC is located, and establish a vibration reference table for each vibration mode in the historical database to store the preset excitation frequency and theoretical response curve; When the graphic control system is in a non-critical operation period, send an active vibration instruction to the vibration trigger and obtain the sampling data of the target PLC during the vibration period; After comparing the sampling data with the theoretical response curve, if the deviation between the two is greater than the threshold, mark the active vibration detection as abnormal in the health record table corresponding to the target PLC.

6. The industrial safety alarm event recording and tracing method based on a graphic control system according to claim 5, characterized in that It also includes: Set a vibration basis function library in the sparse reconstruction algorithm, which is used to store the standard waveform basis functions corresponding to each excitation mode in the vibration reference table; In response to the active vibration detection abnormal mark of the target PLC being written, apply a discount coefficient to the local basis function used by the target PLC in the reconstruction calculation, and increase the priority of the basis functions of the associated PLCs in the multi-source dictionary management module; After completing the reconstruction, write the reference information of the vibration basis function library and the active vibration detection abnormal mark of the target PLC into the historical database.

7. The industrial safety alarm event recording and tracing method based on the graphic control system according to claim 6, characterized in that, It also includes: Set a vibration excitation period configuration item in the database or configuration file to periodically trigger the vibration trigger and collect the vibration response data of the target PLC; In response to detecting that the amplitude or frequency of the vibration response of the target PLC deviates from the theoretical response curve more than the preset range multiple times within the excitation period, automatically reduce the health value of the target PLC below a predetermined threshold; After the health value is reduced, when performing cross-domain reconstruction on the compressed alarm signal uploaded by the target PLC based on the sparse reconstruction algorithm, call the PLC basis functions of other health values greater than or equal to the predetermined threshold to compensate for the data distortion of the target PLC, and write the final reconstruction result into the historical database.

Images